lustre/osc/osc_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Implementation of cl_page for OSC layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39 #include <lustre_osc.h>
  40
  41 #include "osc_internal.h"
  42
  43 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  44 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  45 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
  46                          struct osc_page *opg);
  47
  48 /** \addtogroup osc
  49  *  @{
  50  */
  51
  52 /*
  53  * Page operations.
  54  */
  55 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
  56 {
  57         struct cl_page *page = opg->ops_cl.cpl_page;
  58
  59         LASSERT(!opg->ops_transfer_pinned);
  60         cl_page_get(page);
  61         lu_ref_add_atomic(&page->cp_reference, label, page);
  62         opg->ops_transfer_pinned = 1;
  63 }
  64
  65 static void osc_page_transfer_put(const struct lu_env *env,
  66                                   struct osc_page *opg)
  67 {
  68         struct cl_page *page = opg->ops_cl.cpl_page;
  69
  70         if (opg->ops_transfer_pinned) {
  71                 opg->ops_transfer_pinned = 0;
  72                 lu_ref_del(&page->cp_reference, "transfer", page);
  73                 cl_page_put(env, page);
  74         }
  75 }
  76
  77 /**
  78  * This is called once for every page when it is submitted for a transfer
  79  * either opportunistic (osc_page_cache_add()), or immediate
  80  * (osc_page_submit()).
  81  */
  82 static void osc_page_transfer_add(const struct lu_env *env,
  83                                   struct osc_page *opg, enum cl_req_type crt)
  84 {
  85         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
  86
  87         /* ops_lru and ops_inflight share the same field, so take it from LRU
  88          * first and then use it as inflight. */
  89         osc_lru_use(osc_cli(obj), opg);
  90 }
  91
  92 int osc_page_cache_add(const struct lu_env *env,
  93                         const struct cl_page_slice *slice, struct cl_io *io)
  94 {
  95         struct osc_page *opg = cl2osc_page(slice);
  96         int result;
  97         ENTRY;
  98
  99         osc_page_transfer_get(opg, "transfer\0cache");
 100         result = osc_queue_async_io(env, io, opg);
 101         if (result != 0)
 102                 osc_page_transfer_put(env, opg);
 103         else
 104                 osc_page_transfer_add(env, opg, CRT_WRITE);
 105
 106         RETURN(result);
 107 }
 108
 109 void osc_index2policy(union ldlm_policy_data *policy,
 110                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 111 {
 112         memset(policy, 0, sizeof *policy);
 113         policy->l_extent.start = cl_offset(obj, start);
 114         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 115 }
 116
 117 static const char *osc_list(struct list_head *head)
 118 {
 119         return list_empty(head) ? "-" : "+";
 120 }
 121
 122 static inline s64 osc_submit_duration(struct osc_page *opg)
 123 {
 124         if (ktime_to_ns(opg->ops_submit_time) == 0)
 125                 return 0;
 126
 127         return ktime_ms_delta(ktime_get(), opg->ops_submit_time);
 128 }
 129
 130 static int osc_page_print(const struct lu_env *env,
 131                           const struct cl_page_slice *slice,
 132                           void *cookie, lu_printer_t printer)
 133 {
 134         struct osc_page       *opg = cl2osc_page(slice);
 135         struct osc_async_page *oap = &opg->ops_oap;
 136         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 137         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 138
 139         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 140                           "1< %#x %d %u %s %s > "
 141                           "2< %lld %u %u %#x %#x | %p %p %p > "
 142                           "3< %d %lld %d > "
 143                           "4< %d %d %d %lu %s | %s %s %s %s > "
 144                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 145                           opg, osc_index(opg),
 146                           /* 1 */
 147                           oap->oap_magic, oap->oap_cmd,
 148                           oap->oap_interrupted,
 149                           osc_list(&oap->oap_pending_item),
 150                           osc_list(&oap->oap_rpc_item),
 151                           /* 2 */
 152                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 153                           oap->oap_async_flags, oap->oap_brw_flags,
 154                           oap->oap_request, oap->oap_cli, obj,
 155                           /* 3 */
 156                           opg->ops_transfer_pinned,
 157                           osc_submit_duration(opg), opg->ops_srvlock,
 158                           /* 4 */
 159                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 160                           cli->cl_max_rpcs_in_flight,
 161                           cli->cl_avail_grant,
 162                           osc_list(&cli->cl_cache_waiters),
 163                           osc_list(&cli->cl_loi_ready_list),
 164                           osc_list(&cli->cl_loi_hp_ready_list),
 165                           osc_list(&cli->cl_loi_write_list),
 166                           osc_list(&cli->cl_loi_read_list),
 167                           /* 5 */
 168                           osc_list(&obj->oo_ready_item),
 169                           osc_list(&obj->oo_hp_ready_item),
 170                           osc_list(&obj->oo_write_item),
 171                           osc_list(&obj->oo_read_item),
 172                           atomic_read(&obj->oo_nr_reads),
 173                           osc_list(&obj->oo_reading_exts),
 174                           atomic_read(&obj->oo_nr_writes),
 175                           osc_list(&obj->oo_hp_exts),
 176                           osc_list(&obj->oo_urgent_exts));
 177 }
 178
 179 static void osc_page_delete(const struct lu_env *env,
 180                             const struct cl_page_slice *slice)
 181 {
 182         struct osc_page   *opg = cl2osc_page(slice);
 183         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 184         int rc;
 185
 186         ENTRY;
 187         CDEBUG(D_TRACE, "%p\n", opg);
 188         osc_page_transfer_put(env, opg);
 189         rc = osc_teardown_async_page(env, obj, opg);
 190         if (rc) {
 191                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 192                               "Trying to teardown failed: %d\n", rc);
 193                 LASSERT(0);
 194         }
 195
 196         osc_lru_del(osc_cli(obj), opg);
 197
 198         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 199                 void *value;
 200
 201                 spin_lock(&obj->oo_tree_lock);
 202                 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
 203                 if (value != NULL)
 204                         --obj->oo_npages;
 205                 spin_unlock(&obj->oo_tree_lock);
 206
 207                 LASSERT(ergo(value != NULL, value == opg));
 208         }
 209
 210         EXIT;
 211 }
 212
 213 static void osc_page_clip(const struct lu_env *env,
 214                           const struct cl_page_slice *slice,
 215                           int from, int to)
 216 {
 217         struct osc_page       *opg = cl2osc_page(slice);
 218         struct osc_async_page *oap = &opg->ops_oap;
 219
 220         opg->ops_from = from;
 221         opg->ops_to   = to;
 222         spin_lock(&oap->oap_lock);
 223         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 224         spin_unlock(&oap->oap_lock);
 225 }
 226
 227 static int osc_page_cancel(const struct lu_env *env,
 228                            const struct cl_page_slice *slice)
 229 {
 230         struct osc_page *opg = cl2osc_page(slice);
 231         int rc = 0;
 232
 233         /* Check if the transferring against this page
 234          * is completed, or not even queued. */
 235         if (opg->ops_transfer_pinned)
 236                 /* FIXME: may not be interrupted.. */
 237                 rc = osc_cancel_async_page(env, opg);
 238         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 239         return rc;
 240 }
 241
 242 static int osc_page_flush(const struct lu_env *env,
 243                           const struct cl_page_slice *slice,
 244                           struct cl_io *io)
 245 {
 246         struct osc_page *opg = cl2osc_page(slice);
 247         int rc = 0;
 248         ENTRY;
 249         rc = osc_flush_async_page(env, io, opg);
 250         RETURN(rc);
 251 }
 252
 253 static void osc_page_touch(const struct lu_env *env,
 254                           const struct cl_page_slice *slice, size_t to)
 255 {
 256         struct osc_page *opg = cl2osc_page(slice);
 257         struct cl_object *obj = opg->ops_cl.cpl_obj;
 258
 259         osc_page_touch_at(env, obj, osc_index(opg), to);
 260 }
 261
 262 static const struct cl_page_operations osc_page_ops = {
 263         .cpo_print         = osc_page_print,
 264         .cpo_delete        = osc_page_delete,
 265         .cpo_clip           = osc_page_clip,
 266         .cpo_cancel         = osc_page_cancel,
 267         .cpo_flush          = osc_page_flush,
 268         .cpo_page_touch    = osc_page_touch,
 269 };
 270
 271 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 272                   struct cl_page *page, pgoff_t index)
 273 {
 274         struct osc_object *osc = cl2osc(obj);
 275         struct osc_page   *opg = cl_object_page_slice(obj, page);
 276         struct osc_io *oio = osc_env_io(env);
 277         int result;
 278
 279         opg->ops_from = 0;
 280         opg->ops_to   = PAGE_SIZE;
 281
 282         INIT_LIST_HEAD(&opg->ops_lru);
 283
 284         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 285                                      cl_offset(obj, index));
 286         if (result != 0)
 287                 return result;
 288
 289         opg->ops_srvlock = osc_io_srvlock(oio);
 290         cl_page_slice_add(page, &opg->ops_cl, obj, index,
 291                           &osc_page_ops);
 292
 293
 294         /* reserve an LRU space for this page */
 295         if (page->cp_type == CPT_CACHEABLE) {
 296                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 297                 if (result == 0) {
 298                         result = radix_tree_preload(GFP_NOFS);
 299                         if (result == 0) {
 300                                 spin_lock(&osc->oo_tree_lock);
 301                                 result = radix_tree_insert(&osc->oo_tree,
 302                                                            index, opg);
 303                                 if (result == 0)
 304                                         ++osc->oo_npages;
 305                                 spin_unlock(&osc->oo_tree_lock);
 306
 307                                 radix_tree_preload_end();
 308                         }
 309                 }
 310         }
 311
 312         return result;
 313 }
 314 EXPORT_SYMBOL(osc_page_init);
 315
 316 /**
 317  * Helper function called by osc_io_submit() for every page in an immediate
 318  * transfer (i.e., transferred synchronously).
 319  */
 320 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 321                      enum cl_req_type crt, int brw_flags)
 322 {
 323         struct osc_async_page *oap = &opg->ops_oap;
 324
 325         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 326                  "magic 0x%x\n", oap, oap->oap_magic);
 327         LASSERT(oap->oap_async_flags & ASYNC_READY);
 328         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 329
 330         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 331         oap->oap_page_off  = opg->ops_from;
 332         oap->oap_count     = opg->ops_to - opg->ops_from;
 333         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 334
 335         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 336                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 337                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 338         }
 339
 340         opg->ops_submit_time = ktime_get();
 341         osc_page_transfer_get(opg, "transfer\0imm");
 342         osc_page_transfer_add(env, opg, crt);
 343 }
 344
 345 /* --------------- LRU page management ------------------ */
 346
 347 /* OSC is a natural place to manage LRU pages as applications are specialized
 348  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 349  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 350  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 351  * for free LRU slots - this will be very bad so the algorithm requires each
 352  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 353  * at any time.
 354  */
 355
 356 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 357
 358 /**
 359  * LRU pages are freed in batch mode. OSC should at least free this
 360  * number of pages to avoid running out of LRU slots.
 361  */
 362 static inline int lru_shrink_min(struct client_obd *cli)
 363 {
 364         return cli->cl_max_pages_per_rpc * 2;
 365 }
 366
 367 /**
 368  * free this number at most otherwise it will take too long time to finsih.
 369  */
 370 static inline int lru_shrink_max(struct client_obd *cli)
 371 {
 372         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 373 }
 374
 375 /**
 376  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 377  * we should free slots aggressively. In this way, slots are freed in a steady
 378  * step to maintain fairness among OSCs.
 379  *
 380  * Return how many LRU pages should be freed.
 381  */
 382 static int osc_cache_too_much(struct client_obd *cli)
 383 {
 384         struct cl_client_cache *cache = cli->cl_cache;
 385         long pages = atomic_long_read(&cli->cl_lru_in_list);
 386         unsigned long budget;
 387
 388         LASSERT(cache != NULL);
 389         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 390
 391         /* if it's going to run out LRU slots, we should free some, but not
 392          * too much to maintain faireness among OSCs. */
 393         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 394                 if (pages >= budget)
 395                         return lru_shrink_max(cli);
 396                 else if (pages >= budget / 2)
 397                         return lru_shrink_min(cli);
 398         } else {
 399                 time64_t duration = ktime_get_real_seconds();
 400                 long timediff;
 401
 402                 /* knock out pages by duration of no IO activity */
 403                 duration -= cli->cl_lru_last_used;
 404                 /*
 405                  * The difference shouldn't be more than 70 years
 406                  * so we can safely case to a long. Round to
 407                  * approximately 1 minute.
 408                  */
 409                 timediff = (long)(duration >> 6);
 410                 if (timediff > 0 && pages >= budget / timediff)
 411                         return lru_shrink_min(cli);
 412         }
 413         return 0;
 414 }
 415
 416 int lru_queue_work(const struct lu_env *env, void *data)
 417 {
 418         struct client_obd *cli = data;
 419         int count;
 420
 421         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 422         count = osc_cache_too_much(cli);
 423         if (count > 0) {
 424                 int rc = osc_lru_shrink(env, cli, count, false);
 425
 426                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 427                        cli_name(cli), rc, count);
 428                 if (rc >= count) {
 429                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 430                         ptlrpcd_queue_work(cli->cl_lru_work);
 431                 }
 432         }
 433
 434         RETURN(0);
 435 }
 436
 437 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 438 {
 439         struct list_head lru = LIST_HEAD_INIT(lru);
 440         struct osc_async_page *oap;
 441         long npages = 0;
 442
 443         list_for_each_entry(oap, plist, oap_pending_item) {
 444                 struct osc_page *opg = oap2osc_page(oap);
 445
 446                 if (!opg->ops_in_lru)
 447                         continue;
 448
 449                 ++npages;
 450                 LASSERT(list_empty(&opg->ops_lru));
 451                 list_add(&opg->ops_lru, &lru);
 452         }
 453
 454         if (npages > 0) {
 455                 spin_lock(&cli->cl_lru_list_lock);
 456                 list_splice_tail(&lru, &cli->cl_lru_list);
 457                 atomic_long_sub(npages, &cli->cl_lru_busy);
 458                 atomic_long_add(npages, &cli->cl_lru_in_list);
 459                 cli->cl_lru_last_used = ktime_get_real_seconds();
 460                 spin_unlock(&cli->cl_lru_list_lock);
 461
 462                 if (waitqueue_active(&osc_lru_waitq))
 463                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 464         }
 465 }
 466
 467 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 468 {
 469         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 470         list_del_init(&opg->ops_lru);
 471         atomic_long_dec(&cli->cl_lru_in_list);
 472 }
 473
 474 /**
 475  * Page is being destroyed. The page may be not in LRU list, if the transfer
 476  * has never finished(error occurred).
 477  */
 478 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 479 {
 480         if (opg->ops_in_lru) {
 481                 spin_lock(&cli->cl_lru_list_lock);
 482                 if (!list_empty(&opg->ops_lru)) {
 483                         __osc_lru_del(cli, opg);
 484                 } else {
 485                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 486                         atomic_long_dec(&cli->cl_lru_busy);
 487                 }
 488                 spin_unlock(&cli->cl_lru_list_lock);
 489
 490                 atomic_long_inc(cli->cl_lru_left);
 491                 /* this is a great place to release more LRU pages if
 492                  * this osc occupies too many LRU pages and kernel is
 493                  * stealing one of them. */
 494                 if (osc_cache_too_much(cli)) {
 495                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 496                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 497                 }
 498                 wake_up(&osc_lru_waitq);
 499         } else {
 500                 LASSERT(list_empty(&opg->ops_lru));
 501         }
 502 }
 503
 504 /**
 505  * Delete page from LRU list for redirty.
 506  */
 507 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 508 {
 509         /* If page is being transferred for the first time,
 510          * ops_lru should be empty */
 511         if (opg->ops_in_lru) {
 512                 spin_lock(&cli->cl_lru_list_lock);
 513                 if (!list_empty(&opg->ops_lru)) {
 514                         __osc_lru_del(cli, opg);
 515                         atomic_long_inc(&cli->cl_lru_busy);
 516                 }
 517                 spin_unlock(&cli->cl_lru_list_lock);
 518         }
 519 }
 520
 521 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 522                                 struct cl_page **pvec, int max_index)
 523 {
 524         int i;
 525
 526         for (i = 0; i < max_index; i++) {
 527                 struct cl_page *page = pvec[i];
 528
 529                 LASSERT(cl_page_is_owned(page, io));
 530                 cl_page_delete(env, page);
 531                 cl_page_discard(env, io, page);
 532                 cl_page_disown(env, io, page);
 533                 cl_page_put(env, page);
 534
 535                 pvec[i] = NULL;
 536         }
 537 }
 538
 539 /**
 540  * Check if a cl_page can be released, i.e, it's not being used.
 541  *
 542  * If unstable account is turned on, bulk transfer may hold one refcount
 543  * for recovery so we need to check vmpage refcount as well; otherwise,
 544  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 545  */
 546 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 547 {
 548         if (cl_page_in_use_noref(page))
 549                 return true;
 550
 551         if (cli->cl_cache->ccc_unstable_check) {
 552                 struct page *vmpage = cl_page_vmpage(page);
 553
 554                 /* vmpage have two known users: cl_page and VM page cache */
 555                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 556                         return true;
 557         }
 558         return false;
 559 }
 560
 561 /**
 562  * Drop @target of pages from LRU at most.
 563  */
 564 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 565                    long target, bool force)
 566 {
 567         struct cl_io *io;
 568         struct cl_object *clobj = NULL;
 569         struct cl_page **pvec;
 570         struct osc_page *opg;
 571         long count = 0;
 572         int maxscan = 0;
 573         int index = 0;
 574         int rc = 0;
 575         ENTRY;
 576
 577         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 578         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 579                 RETURN(0);
 580
 581         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 582                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 583         if (!force) {
 584                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 585                         RETURN(-EBUSY);
 586
 587                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 588                         atomic_dec(&cli->cl_lru_shrinkers);
 589                         RETURN(-EBUSY);
 590                 }
 591         } else {
 592                 atomic_inc(&cli->cl_lru_shrinkers);
 593         }
 594
 595         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 596         io = osc_env_thread_io(env);
 597
 598         spin_lock(&cli->cl_lru_list_lock);
 599         if (force)
 600                 cli->cl_lru_reclaim++;
 601         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 602         while (!list_empty(&cli->cl_lru_list)) {
 603                 struct cl_page *page;
 604                 bool will_free = false;
 605
 606                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 607                         break;
 608
 609                 if (--maxscan < 0)
 610                         break;
 611
 612                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 613                                  ops_lru);
 614                 page = opg->ops_cl.cpl_page;
 615                 if (lru_page_busy(cli, page)) {
 616                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 617                         continue;
 618                 }
 619
 620                 LASSERT(page->cp_obj != NULL);
 621                 if (clobj != page->cp_obj) {
 622                         struct cl_object *tmp = page->cp_obj;
 623
 624                         cl_object_get(tmp);
 625                         spin_unlock(&cli->cl_lru_list_lock);
 626
 627                         if (clobj != NULL) {
 628                                 discard_pagevec(env, io, pvec, index);
 629                                 index = 0;
 630
 631                                 cl_io_fini(env, io);
 632                                 cl_object_put(env, clobj);
 633                                 clobj = NULL;
 634                         }
 635
 636                         clobj = tmp;
 637                         io->ci_obj = clobj;
 638                         io->ci_ignore_layout = 1;
 639                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 640
 641                         spin_lock(&cli->cl_lru_list_lock);
 642
 643                         if (rc != 0)
 644                                 break;
 645
 646                         ++maxscan;
 647                         continue;
 648                 }
 649
 650                 if (cl_page_own_try(env, io, page) == 0) {
 651                         if (!lru_page_busy(cli, page)) {
 652                                 /* remove it from lru list earlier to avoid
 653                                  * lock contention */
 654                                 __osc_lru_del(cli, opg);
 655                                 opg->ops_in_lru = 0; /* will be discarded */
 656
 657                                 cl_page_get(page);
 658                                 will_free = true;
 659                         } else {
 660                                 cl_page_disown(env, io, page);
 661                         }
 662                 }
 663
 664                 if (!will_free) {
 665                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 666                         continue;
 667                 }
 668
 669                 /* Don't discard and free the page with cl_lru_list held */
 670                 pvec[index++] = page;
 671                 if (unlikely(index == OTI_PVEC_SIZE)) {
 672                         spin_unlock(&cli->cl_lru_list_lock);
 673                         discard_pagevec(env, io, pvec, index);
 674                         index = 0;
 675
 676                         spin_lock(&cli->cl_lru_list_lock);
 677                 }
 678
 679                 if (++count >= target)
 680                         break;
 681         }
 682         spin_unlock(&cli->cl_lru_list_lock);
 683
 684         if (clobj != NULL) {
 685                 discard_pagevec(env, io, pvec, index);
 686
 687                 cl_io_fini(env, io);
 688                 cl_object_put(env, clobj);
 689         }
 690
 691         atomic_dec(&cli->cl_lru_shrinkers);
 692         if (count > 0) {
 693                 atomic_long_add(count, cli->cl_lru_left);
 694                 wake_up_all(&osc_lru_waitq);
 695         }
 696         RETURN(count > 0 ? count : rc);
 697 }
 698 EXPORT_SYMBOL(osc_lru_shrink);
 699
 700 /**
 701  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 702  * \@npages of LRU slots. For performance consideration, it's better to drop
 703  * LRU pages in batch. Therefore, the actual number is adjusted at least
 704  * max_pages_per_rpc.
 705  */
 706 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 707 {
 708         struct lu_env *env;
 709         struct cl_client_cache *cache = cli->cl_cache;
 710         int max_scans;
 711         __u16 refcheck;
 712         long rc = 0;
 713         ENTRY;
 714
 715         LASSERT(cache != NULL);
 716
 717         env = cl_env_get(&refcheck);
 718         if (IS_ERR(env))
 719                 RETURN(rc);
 720
 721         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 722         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 723                cli_name(cli), npages);
 724         rc = osc_lru_shrink(env, cli, npages, true);
 725         if (rc >= npages) {
 726                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 727                        cli_name(cli), rc, npages);
 728                 if (osc_cache_too_much(cli) > 0)
 729                         ptlrpcd_queue_work(cli->cl_lru_work);
 730                 GOTO(out, rc);
 731         } else if (rc > 0) {
 732                 npages -= rc;
 733         }
 734
 735         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 736                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 737                 atomic_long_read(&cli->cl_lru_busy), npages);
 738
 739         /* Reclaim LRU slots from other client_obd as it can't free enough
 740          * from its own. This should rarely happen. */
 741         spin_lock(&cache->ccc_lru_lock);
 742         LASSERT(!list_empty(&cache->ccc_lru));
 743
 744         cache->ccc_lru_shrinkers++;
 745         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 746
 747         max_scans = atomic_read(&cache->ccc_users) - 2;
 748         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 749                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 750                                  cl_lru_osc);
 751
 752                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 753                         cli_name(cli), cli,
 754                         atomic_long_read(&cli->cl_lru_in_list),
 755                         atomic_long_read(&cli->cl_lru_busy));
 756
 757                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 758                 if (osc_cache_too_much(cli) > 0) {
 759                         spin_unlock(&cache->ccc_lru_lock);
 760
 761                         rc = osc_lru_shrink(env, cli, npages, true);
 762                         spin_lock(&cache->ccc_lru_lock);
 763                         if (rc >= npages)
 764                                 break;
 765                         if (rc > 0)
 766                                 npages -= rc;
 767                 }
 768         }
 769         spin_unlock(&cache->ccc_lru_lock);
 770
 771 out:
 772         cl_env_put(env, &refcheck);
 773         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 774                 cli_name(cli), cli, rc);
 775         return rc;
 776 }
 777
 778 /**
 779  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 780  *
 781  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 782  * Only in the case that the LRU slots are in extreme shortage, it should
 783  * have reserved enough slots for an IO.
 784  */
 785 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 786                          struct osc_page *opg)
 787 {
 788         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 789         struct osc_io *oio = osc_env_io(env);
 790         int rc = 0;
 791
 792         ENTRY;
 793
 794         if (cli->cl_cache == NULL) /* shall not be in LRU */
 795                 RETURN(0);
 796
 797         if (oio->oi_lru_reserved > 0) {
 798                 --oio->oi_lru_reserved;
 799                 goto out;
 800         }
 801
 802         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 803         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 804                 /* run out of LRU spaces, try to drop some by itself */
 805                 rc = osc_lru_reclaim(cli, 1);
 806                 if (rc < 0)
 807                         break;
 808                 if (rc > 0)
 809                         continue;
 810
 811                 cond_resched();
 812                 rc = l_wait_event(osc_lru_waitq,
 813                                 atomic_long_read(cli->cl_lru_left) > 0,
 814                                 &lwi);
 815                 if (rc < 0)
 816                         break;
 817         }
 818
 819 out:
 820         if (rc >= 0) {
 821                 atomic_long_inc(&cli->cl_lru_busy);
 822                 opg->ops_in_lru = 1;
 823                 rc = 0;
 824         }
 825
 826         RETURN(rc);
 827 }
 828
 829 /**
 830  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 831  *
 832  * The benefit of doing this is to reduce contention against atomic counter
 833  * cl_lru_left by changing it from per-page access to per-IO access.
 834  */
 835 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 836 {
 837         unsigned long reserved = 0;
 838         unsigned long max_pages;
 839         unsigned long c;
 840
 841         /* reserve a full RPC window at most to avoid that a thread accidentally
 842          * consumes too many LRU slots */
 843         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 844         if (npages > max_pages)
 845                 npages = max_pages;
 846
 847         c = atomic_long_read(cli->cl_lru_left);
 848         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 849                 c = atomic_long_read(cli->cl_lru_left);
 850         while (c >= npages) {
 851                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 852                         reserved = npages;
 853                         break;
 854                 }
 855                 c = atomic_long_read(cli->cl_lru_left);
 856         }
 857         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 858                 /* If there aren't enough pages in the per-OSC LRU then
 859                  * wake up the LRU thread to try and clear out space, so
 860                  * we don't block if pages are being dirtied quickly. */
 861                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 862                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 863                        max_pages);
 864                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 865         }
 866
 867         return reserved;
 868 }
 869
 870 /**
 871  * osc_lru_unreserve() is called to unreserve LRU slots.
 872  *
 873  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 874  * reasons such as page already existing or I/O error. Those reserved slots
 875  * should be freed by calling this function.
 876  */
 877 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 878 {
 879         atomic_long_add(npages, cli->cl_lru_left);
 880         wake_up_all(&osc_lru_waitq);
 881 }
 882
 883 /**
 884  * Atomic operations are expensive. We accumulate the accounting for the
 885  * same page zone to get better performance.
 886  * In practice this can work pretty good because the pages in the same RPC
 887  * are likely from the same page zone.
 888  */
 889 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 890                                             struct osc_brw_async_args *aa,
 891                                             int factor)
 892 {
 893         int page_count;
 894         void *zone = NULL;
 895         int count = 0;
 896         int i;
 897
 898         if (desc != NULL) {
 899                 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 900                 page_count = desc->bd_iov_count;
 901         } else {
 902                 page_count = aa->aa_page_count;
 903         }
 904
 905         for (i = 0; i < page_count; i++) {
 906                 void *pz;
 907                 if (desc)
 908                         pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
 909                 else
 910                         pz = page_zone(aa->aa_ppga[i]->pg);
 911
 912                 if (likely(pz == zone)) {
 913                         ++count;
 914                         continue;
 915                 }
 916
 917                 if (count > 0) {
 918                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
 919                                             factor * count);
 920                         count = 0;
 921                 }
 922                 zone = pz;
 923                 ++count;
 924         }
 925         if (count > 0)
 926                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
 927 }
 928
 929 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 930                                                 struct osc_brw_async_args *aa)
 931 {
 932         unstable_page_accounting(desc, aa, 1);
 933 }
 934
 935 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 936                                                 struct osc_brw_async_args *aa)
 937 {
 938         unstable_page_accounting(desc, aa, -1);
 939 }
 940
 941 /**
 942  * Performs "unstable" page accounting. This function balances the
 943  * increment operations performed in osc_inc_unstable_pages. It is
 944  * registered as the RPC request callback, and is executed when the
 945  * bulk RPC is committed on the server. Thus at this point, the pages
 946  * involved in the bulk transfer are no longer considered unstable.
 947  *
 948  * If this function is called, the request should have been committed
 949  * or req:rq_unstable must have been set; it implies that the unstable
 950  * statistic have been added.
 951  */
 952 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 953 {
 954         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 955         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 956         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 957         int                      page_count;
 958         long                     unstable_count;
 959
 960         if (desc)
 961                 page_count = desc->bd_iov_count;
 962         else
 963                 page_count = aa->aa_page_count;
 964
 965         LASSERT(page_count >= 0);
 966
 967         dec_unstable_page_accounting(desc, aa);
 968
 969         unstable_count = atomic_long_sub_return(page_count,
 970                                                 &cli->cl_unstable_count);
 971         LASSERT(unstable_count >= 0);
 972
 973         unstable_count = atomic_long_sub_return(page_count,
 974                                            &cli->cl_cache->ccc_unstable_nr);
 975         LASSERT(unstable_count >= 0);
 976         if (unstable_count == 0)
 977                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 978
 979         if (waitqueue_active(&osc_lru_waitq))
 980                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 981 }
 982
 983 /**
 984  * "unstable" page accounting. See: osc_dec_unstable_pages.
 985  */
 986 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 987 {
 988         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 989         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 990         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 991         long                     page_count;
 992
 993         /* No unstable page tracking */
 994         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 995                 return;
 996
 997         if (desc)
 998                 page_count = desc->bd_iov_count;
 999         else
1000                 page_count = aa->aa_page_count;
1001
1002         add_unstable_page_accounting(desc, aa);
1003         atomic_long_add(page_count, &cli->cl_unstable_count);
1004         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1005
1006         /* If the request has already been committed (i.e. brw_commit
1007          * called via rq_commit_cb), we need to undo the unstable page
1008          * increments we just performed because rq_commit_cb wont be
1009          * called again. */
1010         spin_lock(&req->rq_lock);
1011         if (unlikely(req->rq_committed)) {
1012                 spin_unlock(&req->rq_lock);
1013
1014                 osc_dec_unstable_pages(req);
1015         } else {
1016                 req->rq_unstable = 1;
1017                 spin_unlock(&req->rq_lock);
1018         }
1019 }
1020
1021 /**
1022  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1023  * This function will be called by every BRW RPC so it's critical
1024  * to make this function fast.
1025  */
1026 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1027 {
1028         long unstable_nr, osc_unstable_count;
1029
1030         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1031         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1032                 return false;
1033
1034         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1035         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1036
1037         CDEBUG(D_CACHE,
1038                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1039                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1040
1041         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1042          * has one full RPC window of unstable pages, it's a good chance
1043          * to piggyback a SOFT_SYNC flag.
1044          * Please notice that the OST won't take immediate response for the
1045          * SOFT_SYNC request so active OSCs will have more chance to carry
1046          * the flag, this is reasonable. */
1047         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1048                osc_unstable_count > cli->cl_max_pages_per_rpc *
1049                                     cli->cl_max_rpcs_in_flight;
1050 }
1051
1052 /**
1053  * Return how many LRU pages in the cache of all OSC devices
1054  *
1055  * \retval      return # of cached LRU pages times reclaimation tendency
1056  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1057  */
1058 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1059                                      struct shrink_control *sc)
1060 {
1061         struct client_obd *cli;
1062         unsigned long cached = 0;
1063
1064         spin_lock(&osc_shrink_lock);
1065         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1066                 cached += atomic_long_read(&cli->cl_lru_in_list);
1067         spin_unlock(&osc_shrink_lock);
1068
1069         return (cached  * sysctl_vfs_cache_pressure) / 100;
1070 }
1071
1072 /**
1073  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1074  *
1075  * \retval      number of cached LRU pages reclaimed
1076  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1077  *
1078  * Linux kernel will loop calling this shrinker scan routine with
1079  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1080  *
1081  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1082  * to scan and try to reclaim LRU pages, just return 0 and
1083  * osc_cache_shrink_count() will report the LRU page number.
1084  */
1085 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1086                                     struct shrink_control *sc)
1087 {
1088         struct client_obd *cli;
1089         struct client_obd *stop_anchor = NULL;
1090         struct lu_env *env;
1091         long shrank = 0;
1092         int rc;
1093         __u16 refcheck;
1094
1095         if (sc->nr_to_scan == 0)
1096                 return 0;
1097
1098         if (!(sc->gfp_mask & __GFP_FS))
1099                 return SHRINK_STOP;
1100
1101         env = cl_env_get(&refcheck);
1102         if (IS_ERR(env))
1103                 return SHRINK_STOP;
1104
1105         spin_lock(&osc_shrink_lock);
1106         while (!list_empty(&osc_shrink_list)) {
1107                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1108                                  cl_shrink_list);
1109
1110                 if (stop_anchor == NULL)
1111                         stop_anchor = cli;
1112                 else if (cli == stop_anchor)
1113                         break;
1114
1115                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1116                 spin_unlock(&osc_shrink_lock);
1117
1118                 /* shrink no more than max_pages_per_rpc for an OSC */
1119                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1120                                     cli->cl_max_pages_per_rpc ?
1121                                     cli->cl_max_pages_per_rpc :
1122                                     sc->nr_to_scan - shrank, true);
1123                 if (rc > 0)
1124                         shrank += rc;
1125
1126                 if (shrank >= sc->nr_to_scan)
1127                         goto out;
1128
1129                 spin_lock(&osc_shrink_lock);
1130         }
1131         spin_unlock(&osc_shrink_lock);
1132
1133 out:
1134         cl_env_put(env, &refcheck);
1135
1136         return shrank;
1137 }
1138
1139 /** @} osc */